Lymphovascular invasion (LVI) is a critical prognostic marker in urothelial carcinoma of the bladder (UCB), associated with increased tumour aggressiveness, recurrence and lower survival rates. Conventionally, the presence of LVI is confirmed through immunohistochemical (IHC) staining, a method that, while accurate, is both time-consuming and costly. In many clinical environments, especially those with limited resources or in time-sensitive situations, reliance on IHC is impractical. With the increasing integration of artificial intelligence in medicine, there is growing interest in developing non-invasive tools that support earlier diagnosis and more precise treatment planning. A recent study introduced a deep learning-based model designed to preoperatively predict the LVI status of UCB using standard CT imaging, offering clinicians a practical and efficient decision-support tool.
Building a Deep Learning Model from CT Imaging
To develop this predictive model, researchers retrospectively gathered CT scans and clinical data from 577 UCB patients treated across four different medical institutions. The dataset was divided into training, validation and testing cohorts, ensuring that the model's performance could be evaluated on diverse data sources. Radiologists segmented the tumour regions from CT scans, focusing on the largest tumour slice from each of the transverse, sagittal and coronal planes. These slices were standardised and used as input for several convolutional neural networks (CNNs), including ResNet50, InceptionV3, DenseNet121 and VGG11.
Among the CNNs evaluated, ResNet50 exhibited the best overall performance. Its architecture, which includes residual connections to manage deeper network layers, allowed it to extract more nuanced and representative features. These features were captured from the final average pooling layers and visualised using Gradient-weighted Class Activation Mapping (Grad-CAM) to assess model focus and interpretability. Grad-CAM confirmed that the networks primarily attended to tumour margins and internal regions.
Feature extraction was followed by dimensionality reduction using Principal Component Analysis, normalisation and oversampling via the Synthetic Minority Over-sampling Technique (SMOTE) to address class imbalance. The processed features were then used to train three different machine learning models: Decision Tree, XGBoost and LightGBM. Their predictions were combined using a stacking ensemble method, with logistic regression as the meta-learner, resulting in the final deep learning model.
Combining Clinical Data for Enhanced Accuracy
Alongside the deep learning model, the study also considered clinical parameters such as tumour stage, grade, location and growth pattern. Through univariate and multivariate logistic regression, three independent predictors of LVI were identified: pathological T stage, tumour location and tumour growth style. These were used to construct a separate clinical model using logistic regression.
Must Read: Advanced CT Imaging for Bladder Cancer Grading
While the clinical model performed reasonably well on the validation and test datasets, its predictive accuracy was slightly lower than that of the deep learning model. To improve predictive performance further, researchers developed a combined model integrating both deep learning scores and selected clinical variables. This approach not only maintained high accuracy but also demonstrated more consistent performance across internal and external datasets. In the validation set, the combined model achieved an AUC of 0.794 and in the testing set, an AUC of 0.767, indicating robust generalisability and clinical utility.
Clinical Implications and Practical Benefits
The ability to predict LVI status preoperatively has significant implications for treatment planning in bladder cancer. LVI is associated with a greater likelihood of lymph node metastasis and may influence decisions around lymphadenectomy extent, intravesical therapy and systemic treatment. A model capable of flagging LVI-positive cases ahead of surgery can help guide clinicians towards more aggressive surgical strategies or closer postoperative monitoring.
Compared to traditional radiomics approaches, which require manual feature selection and are often subject to operator bias, the use of deep learning offers improved accuracy, efficiency and objectivity. The model developed in this study uses routine CT imaging, a widely available diagnostic tool, making it particularly suitable for use in lower-resource settings. Additionally, as the model does not rely on invasive sampling or complex laboratory processing, it can be easily integrated into existing radiology workflows.
The study also demonstrated that model performance improved with increased network depth, with ResNet50 outperforming ResNet34 and ResNet18. This supports the notion that deeper networks can capture more complex features from medical images, thereby improving prediction accuracy. The use of Grad-CAM further enhances the model's transparency, increasing confidence in its outputs among clinicians.
The study presented a deep learning-based model that predicts lymphovascular invasion in bladder cancer patients using preoperative CT scans. By combining advanced feature extraction from medical images with relevant clinical data, the resulting model offers a non-invasive, accurate and generalisable tool for early risk stratification. It holds particular promise for improving surgical planning and postoperative care in UCB, especially in settings where access to histopathology is limited. Models such as this highlight the growing role of data-driven approaches in personalised and precision oncology.
Source: Insights into Imaging
Image Credit: iStock
